Static Power Reduction in Caches Using Deterministic Naps

Abstract

The dNap architecture is able to accurately transition cache lines to full power state before an access to them. This ensures that there are no additional delays due to waking up drowsy lines. Only cache lines that are determined by the DMC to be accessed in the immediate future are fully powered while others are put in drowsy mode. As a result, we are able to significantly reduce leakage power with no cache performance degradation and minimal hardware overhead, especially at higher associativities. Up to 92% static/Leakage power savings are accomplished with minimal hardware overhead and no performance tradeoff.

Description

CLAIM OF PRIORITY

This application claims priority under 35 U.S.C. 119(e)(1) to U.S. Provisional Application No. 61/983,216 filed 23 Apr. 2014.

TECHNICAL FIELD OF THE INVENTION

The technical field of this invention is cache memory for data processors.

BACKGROUND OF THE INVENTION

Phased caches were previously introduced as cache architecture to reduce the redundant and high-energy consumption caused by reading all data ways on every cache access even though only one of them will be used if the access hits the cache. Phased caches do not query the data arrays in the first cycle of access but rather, wait until a hit is determined before accessing the specific data way hit. This saves dynamic read energy but static energy consumption is not reduced since both the tag and data arrays are ON throughout the program execution.

The rapid increase in microprocessor speed has exceeded the rate of improvement in DRAM (Dynamic Random Access Memory) speed in recent years. This widening performance gap between processors and memories has created several challenges for computer designers since memory performance can easily become a bottleneck to overall system performance. Specifically, processor performance has been observed to increase at about 60% annually, while memory systems lag significantly behind at about 10% annual improvement. To solve this problem, designers turn to memory performance improvements which ultimately dictate the performance and power consumption of processors.

Caching is a common approach used to achieve memory system speed up, by storing data that has been recently used in faster memory. Therefore, using a larger cache could increase the access hit rate, which in turn improves processor speed but this comes at a cost—increased hardware and higher static and dynamic energy consumption.

As a result, there is usually a trade-off between energy and performance in memory system design, since not all accessed memory locations can be stored in faster memories such as caches. Current memory systems designed with SRAMs, DRAMs and/or CAMs, have not been able to catch up with processor performance. As a result, larger caches are often employed in memory systems to bridge this memory processor performance gap. While these large caches offer improved performance, they also increase the power consumed by the processor. An alternative to improve performance is associativity, but it also leads to increased power consumption due to parallel querying of multiple tags. This increasing cache power consumption resulting from the drive for improved performance, cannot be overlooked because caches contribute a significant fraction of the overall power consumed by modern processors. Several authors have concluded that cache/memory systems contribute 30-60% of the total power consumed by processors.

Reducing cache size in an attempt to save power is not an option either, because it leads to higher miss rates and effectively more energy consumption. As a result, several attempts have been made to reduce voltages and design lower power circuits to reduce the high proportion of power consumed by caches/memory systems. However, these circuit level techniques have not been very successful; rather, power dissipation levels have steadily increased with each new microprocessor generation, leading to a renewed interest in architectural approaches that reduce the switching capacitive power component of memory systems without sacrificing performance. In an attempt to save power, some researchers have directed their architectural improvements at better performance because of the observation that improved performance (i.e. less misses) usually lead to less power consumption. Others focus on power reduction techniques targeted at specific aspects of the architecture, with some trade off in performance.

SUMMARY OF THE INVENTION

This invention deterministically powers ON only data RAM lines or group of lines that will be accessed in the immediate future while keeping all other lines powered down. The tag RAMs remain ON to avoid any extra latency associated with powering on a tag RAM/line that is to be accessed. The data RAM on the other hand is deterministically powered ON before access with no extra latency. This is possible in phased caches because hit determination takes a minimum of 2-cycles before data RAM access. Therefore, the power-ON sequence for a set/set-group is triggered on every access to the set/group. Once the hit/miss is determined in the 2^nd cycle, all ways of the set will be ON, then we begin to power down all other ways except the matched (hit/miss) way. The powered ON set/ways are kept ON until the request has been completely processed. All outstanding accesses, in all pipe stages and buffers contribute to the overall power ON state of an individual cache line or group of lines they belong. When way information becomes available, all other ways not scheduled to be read/written are also powered down if no other member of that set or power group of sets needs the way ON.

BRIEF DESCRIPTION OF THE DRAWINGS

These and other aspects of this invention are illustrated in the drawings, in which:

FIG. 1 shows a memory controller module;

FIG. 2 shows implementation in a high performance 2-pipe stage architecture;

FIG. 3 shows implementation in a low power 3 pipe stage architecture;

FIG. 4 shows a cache organized with power groups;

FIG. 5 shows leakage power savings in an L1I cache;

FIG. 6 shows leakage power savings in an L1D cache;

FIG. 7 shows leakage power savings in an L1D cache;

FIG. 8 shows static power savings in an L1I cache;

FIG. 9 shows static power savings in a 4 core processor.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

The deterministic napping technique shown in this invention reduces static/leakage power in caches by leveraging the ability to retain memory contents at low power states. This technique also takes advantage of the fact that data RAMs do not have to be read in the first cycle of cache access while the lines of the referenced set are being transitioned to full power state. These data RAM accesses can occur after tag RAM reads during hit/miss determination or even a cycle after as in phased cache architectures. Unlike conventional drowsy caches, which keep most lines of the data RAM in a low power state, and only restores full power when an access occurs to such low powered lines, the dNap architecture, maintains cache lines that will be accessed in the immediate future, in a fully powered state. This ensures accesses are never stalled while a wake up is being triggered. As a result, dNap caches do not suffer from the performance degradation incurred by conventional drowsy caches, due to accesses to a low powered line. The proposed approach is specifically focused on deterministic naps in only the data RAMs, for two main reasons. First, data RAMs are known to be much larger than the tag RAMs, therefore, they contribute a major portion of static energy. Second, cache accesses are non-deterministic and can occur at any time, starting with a tag RAM read. Therefore, the tag RAMs are always fully powered to avoid delays due to waking a napping tag line.

A Memory Nap Controller (MNC) is used to track in-flight cache transition fully powered lines to a low power napping state. The full power transition is always completed ahead of data RAM access with no extra latency incurred at the time of access. This is enabled by delaying data RAM accesses by 1 or 2 cycles after tag RAM read depending on architecture pipe-line. All current and outstanding accesses, in all pipeline stages and buffers contribute to the overall power ON state of any individual cache line. FIG. 1 shows how S×2^S decoders are used to determine the accessed sets in different cache stages (where S is bit-width of the address' “Set” field) and the contribution of each stage to the final power enable (PE) of each set. The n pipe-stages 101 and m buffers 102 shown in FIG. 1 typically vary by architecture. The S set bits 103 of each pipe stage, are decoded in decoders 104 to activate an L=2^S bit, one hot value representing the specific set decoded. All bits representing a set across the cache pipe-stages and buffers are ORed in gates 104 and 105 to determine which set(s) must be fully powered. For example, all bit-Os of all decoder outputs are fed into OR0 in FIG. 1 to control set 0 power (signal PEO 106). A nap is initiated when a PE transitions from 1 to 0, while wake up is triggered on a transition from 0 to 1. These transitions complete within 1 clock cycle even when waking a larger 512 KB cache. Further static power savings are achievable by transitioning all other ways to a drowsy state after hit/miss determination has chosen a single cache way. If a new access occurs to a set that is attempting to transition unaccessed ways to drowsy state, the new access takes precedence and keeps the line fully powered. The fully powered cache lines are kept in the state until the request has been completely processed. Multiple cache lines can make a dNap power group 107 to reduce the MNC hardware overhead while taking advantage of the built-in low power feature available in custom SRAM memory arrays. For example, power groups of eight, controlled by power group enables (PGEs) as seen in FIG. 1 can be configured. This eliminates the need to use the (log₂ w) LSBs of the set field in the controller logic, thereby reducing the decoder sizes and the number of OR gates significantly, leading to reduced MNC hardware overhead.

In this section, the ease of integrating the dNap architecture with existing cache architectures is discussed as it relates to both high performance and low power architectures. First, FIG. 2 shows how the dNap cache integrates into existing high performance architectures. On a new cache access, the set field 201 of address is decoded in 202, followed by tag RAM 203 reading the first pipe-stage 204 while data RAM reads are delayed to the next pipe-stage 205. Wake up from nap state is also triggered immediately after set decode while tag RAMs are being read. Second, FIG. 3 shows how the dNap cache integrates into existing low power phased cache architecture. Wake up is triggered after set decode in 301 as in their high performance counterparts, the only difference being this low power architecture allows up to 2 cycles 302 and 303 to complete wake-up.

Deterministic napping at the individual cache line allows the maximum number data RAM lines to be kept in nap state, given individual cache line power can now be controlled independently. But this may not be easily achievable in some existing cache architectures which already use memory arrays that group multiple cache lines together for power and area savings. This is usually the case with vendor supplied memory arrays.

To enable the use of existing low power retention available in vendor memories, and to make the deterministic nap architecture more robust, deterministic napping is extended to contiguous cache line groups. The choice of contiguous cache lines is due to the spatial locality of access which suggests that the next contiguous line will most likely be accessed after the current one. Therefore, keeping these contiguous cache lines in the same power state benefits from the proposed deterministic napping scheme by reducing the triggers to transition between nap and full power states by the dNap power controller. For example, FIG. 4 shows an example of grouping eight cache lines 401 to single power group 402; this requires additional OR level for OR′ing the previously OR reduced outputs of sets0 through 7 as seen in FIG. 1.

Vendor supplied memories that already have a built-in low power retention state benefit more from this scheme because they do not need any extra nap state logic per cache line. The trade-off, on the other hand, is possible reduction in static power savings due to more cache lines effectively being fully powered as a result of the power groupings. For example, suppose we have a 2-Way set associative 8 KB cache with 32-byte lines, this cache would have 64 sets per way as seen in FIG. 4. Now, suppose we create groups of 8 cache lines each such that there are 8 power groups per way. Then, a new access to set 9 would trigger full power to all ways of the 2^nd group since the way requested is not known during the first cache pipe stage i.e., tag RAM access. Two cycles later, after hit/miss is determined, the hit way to be read from, or miss way to be allocated, is kept in full power state while the other ways are returned to a nap state if there is no new or pending access to the set and way group. The proposed dNap power groups do not span more than a cache way to enable maximum static power savings after hit/miss determination.

Power-performance trade off of deterministic napping at the individual cache line allows the maximum number of data RAM lines to be kept in Drowsy state, given individual cache line power can now be controlled independently. This offers the best static power savings possible in this architecture because only cache lines offsets to be accessed in the immediate future are fully powered. But this comes at the expense of extra hardware required to implement both the MNC and the individual nap power logic per cache line. Memory system architects can choose to group multiple cache lines into single memory banks to reduce this hardware overhead as needed. Also, to take advantage of the built-in low power feature available in some vendor supplied SRAM memory arrays, system architects can choose to fully power a memory array whenever there is at least an access to any of the lines of the SRAM array.

This eliminates most of the hardware overhead due to napping and wake-up implementation logic but offers lower static power savings because more cache lines are effectively fully powered. Given there are no readily available tools to evaluate static power consumption by dNap caches, we resolved to using Equation 1 for static power proposed by Butts and Sohi

P_static=V_cc*N*K_design*I_leak  (1)

where: V_cc is the supply voltage (Full power is 1.0 V, drowsy power is 0.3 V): N is the number of transistors; K_design is a design dependent parameter; and I_leak is the leakage current which is technology dependent. Since both N and K_design remain constant in both drowsy and full power state, and we already have the V_cc in these states, we evaluate the Drowsy state leakage current I_leak—a as a function of the full power leakage current I_leak using Equation 2 based on the BSIM3 v3.2 equation for leakage.

$\begin{array}{cc}{I}_{\mathrm{leak}}={\mu }_0C_{\mathrm{ox}}\frac{W}{L}{}^{b\left(V_{\mathrm{dd}}-V_{\mathrm{dd}0}\right)}V_t^2\left(1-{}^{\frac{V_{\mathrm{dd}}}{V_t}}\right){}^{\frac{-V_{\mathrm{th}}-V_{\mathrm{off}}}{n^{*}V_t}}& \left(2\right)\end{array}$

where: μ₀ is the zero bias mobility; C_ox, is gate oxide capacitance per unit area,

$\frac{W}{L}$

is the transistor aspect ratio; e^{b(Vdd−Vdd0)} is the DIBL factor derived from the curve fitting method; V_dd0 is the default supply voltage for the technology (V_dd0 is 1.0 V for 70 nm); V_t is the thermal voltage; V_th is threshold voltage which is also a function of temperature; n is the sub threshold swing coefficient; and V_off is an empirically determined BSIM3 parameter which is also a function of threshold voltage. The quantities μ₀, C_ox,

$\frac{W}{L}$

and V_dd0 are statically defined parameters. The DIBL factor b, sub threshold swing coefficient, n and V_off were derived from the curve fitting method based on the transistor level simulations. We calculate the leakage current in drowsy mode I_leak—d as a function of I_leak as follows, where V_ddd is the drowsy mode voltage which is 0.3 V (i.e. 0.7 V less than V_dd) in our simulations.

$\begin{array}{cc}{I}_{\mathrm{leak\_d}}=I_{\mathrm{leak\_d}}*\frac{I_{\mathrm{leak}}}{I_{\mathrm{leak}}}& \left(3\right)\\ =\frac{I_{\mathrm{leak\_d}}}{I_{\mathrm{leak}}}*I_{\mathrm{leak}}& \left(4\right)\end{array}$

Since μ₀, C_ox,

$\frac{W}{L}$

and V_ado are static parameters, they cancel out yielding Equation 6.

$\begin{array}{cc}{I}_{\mathrm{leak\_d}}=\frac{{}^{b\left(V_{\mathrm{dd}}-V_{\mathrm{dd}0}\right)}}{{}^{b\left(V_{\mathrm{dd}}-V_{\mathrm{dd}0}\right)}}*\frac{1-{}^{\frac{V_{{\mathrm{dd}}_d}}{V_t}}}{1-{}^{\frac{V_{\mathrm{dd}}}{V_t}}}& \left(5\right)\\ =\frac{{}^{b\left(-0.77+V_{\mathrm{dd}}-V_{\mathrm{dd}0}\right)}}{{}^{b\left(V_{\mathrm{dd}}-V_{\mathrm{dd}0}\right)}}*\frac{1-{}^{\frac{\left(-0.7+V_{\mathrm{dd}}\right)}{V_t}}}{1-{}^{\frac{V_{\mathrm{dd}}}{V_t}}}& \left(6\right)\end{array}$

The thermal voltage V_t is

$\frac{\mathrm{KT}}{q}$

where: K is the Boltzman constant 1.38088*10⁻²³; q is 1.602*10⁻¹⁹; and T is chosen as 350 K rather than the default 300 K in the hot leakage tool to be consistent with Cacti toolset. We retain the default value of empirical parameter for V_dd, b=2.0 for the 70 nm node. Equation 6 therefore yields Equation 7 after substitution.

I_leak—d=0.24659*I_leak  (7)

Equation 7 which is consistent with estimations is integrated into Cacti for drowsy leakage power evaluations.

The static (or leakage) power on the dNap architecture was measured and compared against equivalently configured conventional caches. Simulations were run on 32 KB level 1(L1) caches with one power enable per line (i.e., w=1), n=3 pipe-line stages and m=4 buffers, and it is expected that at most N ways (where N ways is set associativity) cache lines will be fully powered due to an access in stage 1 and 2, while only 1 cache line in stage 3 and each of the 4 buffers is fully powered in the presence of an access. This is consistent with simulation results, which show more than 92% leakage power savings using the dNap cache architecture. FIG. 5 shows the leakage power savings in a 32 KB L1 Instruction (L1I) cache with 32-byte cache lines and no power groups, compared to an equivalently sized conventional cache, across multiple cache associativities. The increase in leakage power savings as set associativity increases, is due to fewer sets which lead to less dNap hardware overhead, as set associativity increases. This is summarized in Table 1 for an individual 32 KB cache and Table 2 at the processor core level. These two tables also show that the proposed dNap technique incurs similar hardware overheads in the existing DVS technique at higher associativities. This is because the number of sets reduce as associativity increases, thereby leading to fewer input and outputs on the decoder logic, and effectively, fewer OR gates.

TABLE 1
Hardware Overhead Comparison in 32 KB Cache
Associativity	DVS (%)	dNap (%)
1 Way	+6.93	+13.54
2 Way	+6.61	+10.23
4 Way	+6.07	+8.00
8 Way	+5.26	+6.28
16 Way	+4.15	+4.66

TABLE 2
Hardware Overhead Comparison at Processor Core Level
Associativity	DVS (%)	dNap (%)
1 Way	+1.63	+3.18
2 Way	+1.55	+2.40
4 Way	+1.43	+1.88
8 Way	+1.24	+1.48
16 Way	+0.97	+1.10

Simulation results indicate more than 92% leakage power savings is achievable with the proposed dNap cache architecture. FIG. 5 shows the leakage power savings in a 32 KB L1 Instruction (L1I) cache with 32-byte lines, compared to an equivalently sized conventional cache, for direct mapped through 16 Way set associativity. The increase in leakage power savings as set associativity increases, is due to lower percentage hardware increase due to nap state implementation as set associativity increases as seen in Table 1. This rate of hardware reduction is greater than that of the increase in fully powered lines due to set associativity seen as in simulation results. This explains the consistent increase in leakage power savings with increasing associativity.

The dNap scheme shows a slightly better leakage power savings percentage in the L1 Data cache because there were fewer accesses to the L1D in the 500 million cycle simulation window across the benchmarks. This allows the L1D cache to have a higher proportion of cache lines in nap state during program execution.

The significant static power savings (more than 90%) due to the dNap architecture does not vary much across different associativities, because the number of fully powered cache lines only varies in the first 2 cache pipe-stages before hit/miss way is known. This difference is less than 1% because simulation configurations use 1024 cache lines (i.e., 32 KB cache, 32 byte lines), and the maximum number of extra lines in the 16 Way cache configuration are the 15 extra ways in the first 2 pipe stages before hit/miss determination. This results in only 30 extra cache lines fully powered out of 1024 lines versus the direct mapped cache alternative.

Also, there can only be a maximum of “2*N ways+n−2+m” fully powered lines at any given cycle during program execution in the proposed dNap architecture, where N ways is associativity, n is the number of pipe-stages and m is the number of buffers. This suggests that the performance of the dNap technique will show only negligible variations in static/leakage power savings as reflected in FIGS. 5 and 6. For example, for N ways=16, n= and m=4, the number of fully powered lines in any cycle, varies from none, to a maximum of 37 of the total 1024 cache lines leading to over 90% power savings.

The static power reduction benefits of deterministic napping is also evaluated in low power wearable medical devices. FIG. 7 is focused on the medical applications of SPEC2006 processor benchmark. These include hmmer which is used in computational biology to search for patterns in DNA sequences, libquantum, used in human cryptography and namd, a parallel program used in the simulation of large bio-molecular systems. Povray calculates and simulates the way rays of light travels and hits the human eye, while sphinx is a widely known speech recognition program. The significant dNap static power savings do not vary much across different associativities and applications, because there can only be a maximum of “2*Nways+m+n−2” fully powered lines at any given cycle during program execution, where n and m are the pipe-stage and buffer depths, respectively. This is because tag access and comparison occurs in the first 2 cycles and all cache ways (2*N ways) must be fully powered before a single hit/miss way is determined. Beyond that, accesses in other pipe-line stages and buffer (i.e., m+n−2) keep only the cache hit/miss way oft he referenced set in full power state. This explains the minor variations in leakage power savings across applications.

For example, in a 16-way, 32 KB cache, with 32-byte lines, n=3 and m=4, the number of fully powered lines in any cycle, varies from 0 to a maximum of 37 of the total 1024 cache lines. As a result, the maximum possible variation is 37/1024, which explains the minimal variation in the static power measurements across the different benchmarks and similarly configured L1I and L1D caches. This is unlike dynamic power which significantly varies with applications and access patterns. These light static power variations are partly due to multiple in-flight references to the same set, and partly due to some unoccupied pipe-stages or buffer slots in one cache and/or application but not in the other. Cache architectures with deeper pipe-stages and buffers are expected to show more power savings variation.

The overall leakage power reduction across the cache hierarchy is further evaluated while highlighting the effect of dNap logic and dynamic power due to nap state transitions. This was achieved using the default Intel configuration in the sniper simulator, with 64-byte cache lines, 32 KB L1I and L1D with 4-Way and 8-Way associativity respectively. The L2 and L3 were configured as 8-Way 256 KB and 16-Way 32 MB respectively. The McPAT power measurements are summarized in Table 3. It shows the overhead due to nap state transitions are negligible while dNap power savings are still significant, with the highest power contribution due to the always fully powered dNap logic.

TABLE 3
Total Leakage Power in L1s, L2 and L3
	dNap Cache Power (W)	Conv.	Savings
Benchmarks	Wake	logic	other	(W)	(%)
hmmer	0.00031	1.86	1.00	9.16	68.68
libquantum	0.00006	1.86	0.97	9.16	69.06
namd	0.00014	1.86	0.98	9.16	68.91
povray	0.00017	1.86	0.97	9.16	68.98
sphinx3	0.00014	1.86	0.98	9.16	68.96

Leakage (or static) power reduction due to dNaps was also evaluated in a multi core environment. FIG. 8 shows the total static power reduction in all L1 Instruction caches across a 4-core processor. The result is significant static power reduction which does not exceed those of the DVS technique. This is because the compared flavor of DVS technique waits until access reaches the data RAM before waking up the RAMs; as a result, the DVS is in lower power state for longer periods than the dNap technique. The dNap however, does not incur the performance penalty of the DVS for the static power savings it offers.

FIG. 9 shows a similar trend in static power savings at the 4-core processor level. The only difference being a lower static power reduction, due to other processor components not participating in deterministic napping.

It is worth noting that more cache lines per dNap group leads to fewer wake-up transitions due to more fully powered lines over the course of program execution. It was also observed that all power groups in all benchmarks evaluated in this work, completely transitioned in and out of nap state within a single clock cycle.

Both the Simple scalar toolset and Cacti v6.5 toolset was used as the basis of the simulator development for static power evaluation. While there are multiple flavors of these tools, none completely model the architectural technique shown in this invention. Therefore, a robust simulator was developed using both existing tools as basis. The state of all cache lines are tracked per cycle and the static power for each line is computed using Equations 1 and 7. The total static energy for 500 million cycles of simulation was collected for different 32 KB cache configurations on SPEC2006 benchmarks and compared with conventional non-drowsy caches. Table 4 gives a brief summary of the default configurations used across all of the simulations.

TABLE 4
Simulation Configuration
Parameters                   Value
Instruction Fetch Queue Size 4 Instructions
Instruction Decode Width     4 Instructions Per Cycle
Instruction Issue            4 Instructions Per Cycle
L1 Instruction Cache Latency 3 cycles
L1 Data Cache Latency        3 cycles
L2 Unified Cache Latency     11 cycles
Main Memory Latency          26 cycles
TLB Miss Latency             30 cycles
Memory Access Width          32 Bytes

Claims

1. A cache memory system, comprising:

a cache memory operable to store data in a plurality of locations defined by a plurality of addresses and divided into a plurality of cache ways;

a memory nap controller operable to track in-flight cache accesses and further operable to control the desired power state of a plurality of data RAM lines;

a processing block operable to group a plurality of contiguous cache lines into a single power group.

2. The cache memory system of claim 1, wherein:

the power state of said data RAM lines is determined by the memory nap controller based on all current an outstanding data RAM accesses in all pipeline stages and buffers.

3. The cache memory system of claim 1, wherein:

the memory nap controller is operable to transition the data RAM lines to a fully powered operating state or to a low power nap state.

4. The cache memory system of claim 1, wherein:

the memory nap controller is operable to complete the transition of the data RAM lines to a fully powered state ahead of the actual data RAM access without any additional latency incurred at the time of the access.

5. The cache memory system of claim 1, wherein:

the memory nap controller is operable to place the cache lines that will not be used in the immediate future in a low power state, and keep only the cache lines determined to be accessed in the immediate future in a fully powered on state.

6. The cache memory system of claim 1, wherein:

the memory nap controller is operable, after a hit/miss determination has selected a single cache way to place all other ways in a low power state.

7. The cache memory system of claim 6, wherein:

the memory nap controller will, in the case of a new access to a cache way that is transitioning to a low power state maintain said cache way in a powered up state until the access has completed.